Ensemble deep learning model for leak detection from multi-channel acoustic signals

Mikami, Nao; Aizawa, Kosuke ; Ueki, Yoshitaka*; Michel, F.*; Fache, J.*

The present study evaluates the basic feasibility of an ensemble deep learning model to detect leakage from multi-channel acoustic signals in a steam generator (SG) of a sodium-cooled fast reactor (SFR). The acoustic signals from the bubbling and the gas blowout are measured by an array sensor in a basic experimental apparatus for SG to simulate noise and leak signals. Time-frequency representations (TFRs) are produced from these acoustic signals as the inputs of convolutional neural networks (CNNs). Three typical CNNs are introduced as candidates for the base model of ensemble deep learning. The proposed ensemble deep learning model reaches an accuracy of 95.43%, improved by 4.90% from the solo deep learning model. This result indicates that the proposed ensemble deep learning model has the potential to detect leakage more precisely in an actual SG of SFR.